How should continuous clinical variables be categorized in high-dimensional prediction models? A boosting deselection approach

This repository implements a methodology for data-driven cut-point selection in high-dimensional settings. The approach combines generalized additive models (GAMs), component-wise gradient boosting, and a deselection mechanism to jointly perform variable selection and identify clinically meaningful cut-points for continuous predictors.

The method consists of four main steps:

1. Estimate smooth predictor effects using generalized additive models (GAMs).
2. Generate candidate cut-point configurations from step-function approximations.
3. Perform component-wise gradient boosting using all candidate discretizations as base-learners.
4. Select the optimal cut-point configuration for each predictor based on the risk reduction.